Ask MishMash

Ask a question about the centre and get an answer written from this website's own pages. The answer is generated by an AI model, which can be wrong or leave things out, so every answer names the pages it came from and those pages are the record.

The model runs in your browser through WebLLM. Your question and the answer stay on your machine and are not sent anywhere. Two things are fetched from elsewhere the first time: the model itself, from its public host, and the runtime, from a package mirror. After that the page works offline. The passages come from this site: pages, news, events, and the directory of people, projects and institutions.

What it can and cannot do

It can answer from what this site publishes: the about pages, the work packages, news, events, and the directory of people, projects and institutions, in English and in Bokmål. Ask it what a work package studies, what the terms of use allow, when something happened, or who works on a topic, and it will answer from those pages and link them.

It cannot tell you anything the site does not say. It has no access to e-mail, internal documents, the funding application or anything behind a login, and it does not know what happened after the last time the site was built. It cannot book a room, send a message or fill in a form.

It can also be wrong. A small model writing from a handful of passages may miss a qualification, merge two facts or state something with more confidence than the source does. The pages it names are the record; the answer is a way of finding them. A question in Norwegian is answered from the Norwegian pages where they exist, and there are fewer of those than English ones.

How well it finds the right page is measured rather than asserted: a set of questions in both languages is scored at every change, and the result is on the issue. Something wrong or missing is worth telling us at contact@mishmash.no.

Loading knowledge base…

Sizes are approximate GPU memory requirements. The first load downloads the model (several minutes on a slow connection); later loads use the local cache. If your GPU lacks 16-bit shader support, a compatible variant is selected automatically.

Troubleshooting
  • Browser: WebGPU works best in recent Chrome or Edge (113+). Firefox and Safari are adding support; if the page reports that WebGPU is unavailable, try Chrome or Edge first.
  • First load is slow: the model (0.5–2.5 GB) is downloaded once and cached by the browser. Keep the tab open until the progress bar completes.
  • Load fails or the tab crashes: your GPU may not have enough memory for the selected model. Pick a smaller one (Gemma 3 1B or Llama 3.2 1B) and try again.
  • Shader-f16 errors: some GPUs lack 16-bit shader support. The page detects this and automatically loads a 32-bit variant (larger, but compatible). The Debug info below shows what was selected.
  • Redownload a broken model: clear the site's cached data (browser settings → site settings → view permissions and data stored across sites → this site → clear data), then reload the page.
  • Edge on Linux: Edge does not enable Vulkan/WebGPU by default on Linux, even on machines where Chrome works. Enable Vulkan and WebGPU Developer Features at edge://flags, restart Edge, and check edge://gpu shows Vulkan: Enabled. Alternatively launch with microsoft-edge --enable-features=Vulkan.
  • Linux / Ubuntu: WebGPU needs Vulkan.
    1. Install Vulkan drivers: sudo apt install mesa-vulkan-drivers vulkan-tools
      For NVIDIA: install the proprietary driver via Software & Updates → Additional Drivers.
      For Intel Arc: sudo apt install intel-media-va-driver-non-free.
    2. Verify Vulkan works: vulkaninfo | head -5
    3. In Chrome or Edge, enable Vulkan and WebGPU Developer Features at chrome://flags (in Edge: edge://flags).
    4. Enable hardware acceleration: Settings → System → Use hardware acceleration when available.
    5. Check chrome://gpu (in Edge: edge://gpu) and look for Vulkan: Enabled under Graphics Feature Status.
    6. As a last resort, launch the browser with google-chrome --enable-features=Vulkan,UseSkiaRenderer
Debug info 
Collecting diagnostics…